The aim of the research is to check the effectiveness of the method of automatic identification of “toxic” comments of users in social networks based on semantic proximity. The article carries out a linguistic analysis of examples of “toxic” behavior, defines the criteria of “toxicity” and the main lexical and stylistic features of “toxic” texts. The analysis of the latest works on the topic gives a general idea of the current methods of identifying “toxicity”. A solution for identifying “toxic” comments based on the idea of the lack of semantic proximity between the text of the post and the “toxic” comment is tested. The scientific novelty lies in the fact that the work proposes for the first time to use the criterion of semantic proximity to identify “toxic” comments, which is a fairly simple and effective solution. Moreover, such studies have not been conducted earlier within the framework of the most popular Russian-language social network VKontakte. As a result of the research, it was found that determining the semantic proximity between a post and a comment is a fairly effective way to determine the relevance of a comment and, consequently, its probable “toxic” connotation. It was also found that the cosine similarity metric is suitable for conducting experiments to identify “toxicity”, but to improve the results, it can be supplemented with other machine learning methods.